Pulmonary Nodule Detection Based on Multi-branch 3D Squeeze-and-Excitation Network

Yuechao Zhang, Jianxin Zhang, Chao Che, Kai Lin, Dongsheng Zhou, Qiang Zhang, Xiaopeng Wei

Abstract: Convolutional neural networks based regional proposal networks (RPN) have recently achieved breakthrough results in a variety of medical image detection tasks. Among them, 3D RPN network in conjunction with advanced ResNet gains the state-of-the-art performance. However, current 3D RPN network cannot fully consider different hierarchically spatial information and relationship between channels contained in networks themselves. To overcome this problem, this work proposes a novel multi-branch 3D Squeeze-and-Excitation (SE) network, i.e., Mb3DSENet, which embeds SE-Inception and SE-Residual-Inception blocks into existing 3D RPN network by appropriately assembling SE unit, Inception module and residual connection. Mb3DSENet not only inherits the merits of the basic 3D RPN network, but also enhances its spatial representation power and selectively highlights channel-wise feature responses. Experiment results on the public LUNA16 dataset demonstrate the effectiveness of the proposed network.